---
license: mit
task_categories:
- fill-mask
- text-generation
language:
- en
tags:
- art
pretty_name: Fiction 1B
size_categories:
- 1B<n<10B
---

# Fiction 1B

More than 1B words of narrative fiction sourced from [Project Gutenberg](https://www.gutenberg.org/), [AO3](https://archiveofourown.org/), and [Internet Archive](https://archive.org/).

## Dataset Details

### Dataset Description

This contains the text of roughly 20,000 works of narrative fiction from the above sources.
From the original full texts, a [genre classifier](https://huggingface.co/classla/xlm-roberta-base-multilingual-text-genre-classifier) was applied at the paragraph level to remove license text, metadata, and other content suspected not to be narrative prose.

#### Misc

- **Curated by:** Shawn Rushefsky - [🤗](https://huggingface.co/shawnrushefsky) | [github](https://github.com/shawnrushefsky)
- **Funded by:** [Salad Technologies](https://salad.com)
- **Language(s) (NLP):** English
- **License:** MIT

### Dataset Sources

More information about specific source documents can be found in `doc_index.csv`

- Project Gutenberg: 76.4%
- Archive of our Own (AO3): 22.2%
- Internet Archive: 1.4%

## Uses

The dataset is intended to be used for training language models on the syntactic patterns of narrative fiction.

### Direct Use

- Fill-Mask training
- Text Generation training
- Research

### Out-of-Scope Use

- Applications outside of fiction

## Dataset Structure

`data.zip` contains a CSV file where each row contains the source, a document ID, paragraph index, approximately 500 words of text, and a word count for that section.

## Dataset Creation

### Curation Rationale

While much of this content is already present in extremely large web-scaped datasets, there is a scarcity of more approachable medium-sized datasets that focus specifically on narrative fiction.
Datasets such as [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) with trillions of tokens are not practical for the average developer to work with.

### Source Data

#### Data Collection and Processing

**Project Gutenberg**

Project Gutenberg hosts a [catalog CSV](https://www.gutenberg.org/cache/epub/feeds/pg_catalog.csv) that includes metadata such as title, author, and subjects.
I filtered based on the presence of fiction-related keywords in the Subjects column, and used a python script to bulk download texts.

```python
fiction_keywords = [
    'fiction', 'novel', 'stories', 'tale', 'adventure',
    'mystery', 'romance', 'fantasy', 'horror', 'detective',
    'science fiction', 'historical fiction', 'western',
    'thriller', 'suspense'
]
```

**AO3**

For AO3, I used the [ao3-api](https://pypi.org/project/ao3-api/) python package to gradually paginate through the archive, filtering to English language work with at least 15,000 words but fewer than 500,000, sorted by “Kudos”, a measure of user favor.

**Internet Archive**

For Internet Archive, I used their search endpoint, and a significant amount of keyword filtering.
Ultimately I did not get much content from this source due to licensing restrictions.

#### Who are the source data producers?

Professional and amateur writers of long-form narrative fiction in the English language over the last few hundred years.

#### Personal and Sensitive Information

This dataset contains only works of fiction.

## Bias, Risks, and Limitations

The source text comes from a diverse set of english-language narrative fiction spanning hundreds of years of authorship, and may include subject matter and phrasing that offend.
The age of much of the material from Project Gutenberg is such that white men from before the civil rights movement are vastly disproportionately represented as authors.
Additionally, contemporary commercial fiction is nearly all but excluded due to licensing restrictions.

### Recommendations

Use at your own risk.